Child of #15: architectural / civil reconstruction path (room scans, BIM) #18

New issue

Closed

opened 2026-06-17 18:25:49 +00:00 by Mark · 0 comments

Mark commented

2026-06-17 18:25:49 +00:00

Owner

Parent

Part of #15 — refine pc2cad strategy. The end-to-end pipeline forks into three reconstruction paths. This is the architectural / civil / BIM path.

What "architectural" means here

Inputs from building / site scans:

Room scans (interior walls, floors, ceilings, doors, windows)
Exterior building scans (facades, roofs, columns, beams)
Topographic / site surveys (terrain, roads, utilities)
BIM-style outputs (IFC, Revit) are the goal, not editable feature CAD

Pipeline characteristics

Input class: point cloud dominated by planar regions (walls, floors, ceilings) with some cylindrical (columns, pipes) and irregular geometry (terrain, vegetation)
Primary primitives: planes (walls, floors, ceilings), boxes (rooms), cylinders (columns, pipes), terrain meshes
Output format: IFC for BIM, optionally 2D plan/elevation drawings derived from the 3D model
Feature detection: semantic segmentation (wall vs floor vs ceiling vs opening) is the goal, not parameter extraction
Scale: much larger than mechanical — entire rooms/buildings, often gigabytes of point data, octree required
Topology: rooms are watertight volumes; buildings are open shells (exterior)

What to design

Semantic segmentation — given a room scan, label each planar region as wall/floor/ceiling/opening. Use vertical/horizontal orientation cues + adjacency to other planes (a plane adjacent to a floor and ceiling on opposite sides is a wall).
Scale handling — point clouds are GBs. Design the octree streaming strategy, downsampling per region, and chunking for memory. The current load_point_cloud reads the whole cloud into memory — won't work for a building.
Output —
- IFC for BIM (Industry Foundation Classes, the open BIM standard)
- Plan/elevation 2D drawings derived by projecting the 3D model (DXF or PDF)
- Coverage report: fraction of input scan covered by extracted surfaces
MVP — a single rectangular room with one door and one window. If we can produce IFC for that, the pipeline handles real interiors.
Library options — for IFC: IfcOpenShell (python bindings exist). For 2D drawings: ezdxf. Both are mature and well-documented.

Scope

In scope: design doc for the architectural path. Defines semantic segmentation rules, scale handling strategy, output formats, MVP scope, library choices, test corpus.
Out of scope: implementing the architectural pipeline. Separate tickets after design approval.

Acceptance

Document at docs/STRATEGY-ARCHITECTURAL.md covering: semantic segmentation rules, scale handling (octree, streaming, chunking), output formats (IFC primary, 2D drawings secondary), MVP scope with specific test room, library selection (IfcOpenShell, ezdxf), test corpus
Identification of the scale-handling risk (this is likely the riskiest single piece — current code reads whole clouds)
Decision recorded in #15 as a link back to this child

Why this matters

A scan-to-CAD tool that only handles mechanical parts misses the largest market segment (architecture / civil / BIM). Without a separate design for the architectural path, the mechanical pipeline will be stretched to fit and produce meaningless results on building scans (every wall becomes a 'hole' because it's a large vertical plane).

## Parent Part of #15 — refine pc2cad strategy. The end-to-end pipeline forks into three reconstruction paths. This is the architectural / civil / BIM path. ## What "architectural" means here Inputs from building / site scans: - Room scans (interior walls, floors, ceilings, doors, windows) - Exterior building scans (facades, roofs, columns, beams) - Topographic / site surveys (terrain, roads, utilities) - BIM-style outputs (IFC, Revit) are the goal, not editable feature CAD ## Pipeline characteristics - **Input class:** point cloud dominated by planar regions (walls, floors, ceilings) with some cylindrical (columns, pipes) and irregular geometry (terrain, vegetation) - **Primary primitives:** planes (walls, floors, ceilings), boxes (rooms), cylinders (columns, pipes), terrain meshes - **Output format:** IFC for BIM, optionally 2D plan/elevation drawings derived from the 3D model - **Feature detection:** semantic segmentation (wall vs floor vs ceiling vs opening) is the goal, not parameter extraction - **Scale:** much larger than mechanical — entire rooms/buildings, often gigabytes of point data, octree required - **Topology:** rooms are watertight volumes; buildings are open shells (exterior) ## What to design 1. **Semantic segmentation** — given a room scan, label each planar region as wall/floor/ceiling/opening. Use vertical/horizontal orientation cues + adjacency to other planes (a plane adjacent to a floor and ceiling on opposite sides is a wall). 2. **Scale handling** — point clouds are GBs. Design the octree streaming strategy, downsampling per region, and chunking for memory. The current `load_point_cloud` reads the whole cloud into memory — won't work for a building. 3. **Output** — - IFC for BIM (Industry Foundation Classes, the open BIM standard) - Plan/elevation 2D drawings derived by projecting the 3D model (DXF or PDF) - Coverage report: fraction of input scan covered by extracted surfaces 4. **MVP** — a single rectangular room with one door and one window. If we can produce IFC for that, the pipeline handles real interiors. 5. **Library options** — for IFC: IfcOpenShell (python bindings exist). For 2D drawings: ezdxf. Both are mature and well-documented. ## Scope - **In scope:** design doc for the architectural path. Defines semantic segmentation rules, scale handling strategy, output formats, MVP scope, library choices, test corpus. - **Out of scope:** implementing the architectural pipeline. Separate tickets after design approval. ## Acceptance - Document at `docs/STRATEGY-ARCHITECTURAL.md` covering: semantic segmentation rules, scale handling (octree, streaming, chunking), output formats (IFC primary, 2D drawings secondary), MVP scope with specific test room, library selection (IfcOpenShell, ezdxf), test corpus - Identification of the scale-handling risk (this is likely the riskiest single piece — current code reads whole clouds) - Decision recorded in #15 as a link back to this child ## Why this matters A scan-to-CAD tool that only handles mechanical parts misses the largest market segment (architecture / civil / BIM). Without a separate design for the architectural path, the mechanical pipeline will be stretched to fit and produce meaningless results on building scans (every wall becomes a 'hole' because it's a large vertical plane).